Somewhere in New York's hedge fund machinery, an artificial intelligence system has been making investment calls since last October. Not generating research reports for human review. Not flagging opportunities for analysts to chase down. Actually producing trading signals—the kind an institutional investor puts money behind.
The system comes from KelAI, a three-person outfit that emerged from Y Combinator this spring with a founder who knows both sides of the problem intimately. Jeremie Cohen spent years managing systematic equity portfolios at WorldQuant and leading machine learning initiatives at Millennium Management before deciding the only way to build AI that works in finance was to strip away the usual venture-backed pretense and focus on what portfolio managers actually need.
Which is to say: alpha. Real, measurable, bet-your-fund-on-it alpha.
The timing carries a certain defiance. Last September—seven months before KelAI's system went live—Citadel's Ken Griffin told Bloomberg flat-out that generative AI "fails to help hedge funds produce alpha." He wasn't wrong to be skeptical. Quantitative shops have been throwing machine learning at markets for the better part of a decade with results that range from promising to catastrophic, and the graveyard of research prototypes that never made it to production trading is crowded.
Cohen's pitch is that most of those efforts missed the point. They built tools. He built an engine.
The Full Loop
KelAI describes itself as "agentic," Silicon Valley's current favorite term for AI that can chain together multiple steps without constant human handholding. In this case, that means the platform plugs into a fund's proprietary data infrastructure, investment mandate, universe constraints, and risk parameters, then runs what the company calls the complete research cycle: generating ideas, pulling and analyzing data, writing and testing code, backtesting strategies, validating results, monitoring performance, and—crucially—learning from portfolio manager feedback.
It's designed around the institutional realities that usually kill fintech products before they get out of the pilot phase. Data rights. Intellectual property firewalls. Risk limits that shift with market conditions. The particular vocabulary and reporting style each PM has developed over years of managing money. Cohen managed systematic equity portfolios at WorldQuant before his stint at Millennium—he knows where fintech idealism typically crashes into trading floor pragmatism.
"AI as the engine itself—built to create edge with compute, not headcount," the company posted on LinkedIn in mid-June, a framing that positions the product less as a research assistant and more as a fundamental reimagining of how alpha gets manufactured.
Whether that's marketing or reality is, of course, the question.
The Founder's Unusual Path
Cohen's background is not the typical fintech founder story. Before WorldQuant and Millennium, he worked in data engineering at SolveBio. Before that, he trained as a nuclear engineer. The combination—quantitative rigor, machine learning chops, hands-on portfolio management experience—is relatively rare. Most startups taking a run at this problem come from either pure tech or pure finance. Cohen has lived in both worlds, which may explain why he's focused on solving workflow integration problems that pure technologists often miss.
WorldQuant itself published a perspective last May acknowledging that agentic AI systems are "moving from implementing to helping generate ideas," though the firm was careful to emphasize that human accountability remains essential. Cohen appears to have taken that observation and pushed it a step further, automating idea generation while keeping portfolio managers firmly in the decision seat.
The Client That Appeared Early
Here's the detail that stands out: KelAI's launch materials claim the system has been generating live signals since October 2025. That timeline suggests Cohen landed an institutional client early in the company's development—an uncommon achievement for a technical infrastructure product in an industry that treats new vendors the way bomb squads treat suspicious packages.
The company hasn't disclosed performance metrics. The launch post describes the output as "valuable alpha in production," but there are no information ratios, Sharpe ratios, hit rates, or drawdown figures on offer. The institutional client remains unnamed, which is standard practice when hedge funds treat their technology stack as proprietary intelligence.
As of June, KelAI is hiring for four roles: Applied AI Engineer, AI Researcher, Quant Developer, and Infrastructure Engineer. For a three-person team already running production signals, those hiring priorities suggest the plan is to scale toward multiple deployments rather than just nursing a single client relationship.
A Crowded Field, A Narrow Claim
KelAI is hardly the first company to take a swing at autonomous or crowd-sourced alpha generation. Numerai Signals runs a community platform feeding an institutional market-neutral fund. QuantConnect's Alpha Streams lets funds license algorithms built on its infrastructure. CrunchDAO crowdsources research signals. S&P Global maintains a library of stock-selection factors that any fund can license.
What distinguishes KelAI—at least in how it presents itself—is the promise of full autonomy and deep integration with a specific fund's workflows, data infrastructure, and operational constraints. Most competing platforms either deliver factors and signals that funds must then integrate themselves, or provide research environments where humans still drive the process end-to-end.
Recent academic work offers both encouragement and warning. A January paper proposed an agentic web-searching framework that reported daily alpha on the Russell 1000, though only at the research stage with all the caveats that implies. A March study modeled how widespread AI adoption in trading could actually accelerate alpha decay through algorithmic homogenization—a pointed reminder that if autonomous alpha generation works, it may not work for long once the entire industry has access to similar tools.
The company hasn't disclosed funding beyond Y Combinator's standard investment. Operating out of New York, it's planting itself at the intersection of the city's hedge fund ecosystem and its expanding AI engineering talent pool—a bet that both worlds are finally ready to meet in the middle.
The Proof Is in the P&L
Whether KelAI's autonomous engine delivers sustained alpha at institutional scale remains an open question. Performance in finance has a way of looking brilliant right up until it doesn't, and a few months of live signals—while more than most startups can claim—hardly constitutes a track record.
But Cohen has done something that matters in an industry where talk is cheap and capital is expensive: he's convinced someone with actual money on the line to run his system in production. Not in a sandbox. Not in a pilot program with monopoly money. In production, where bad calls cost real dollars and explanations to risk committees get uncomfortable fast.
In a business where alpha is measured in basis points and skepticism is the default setting, that early validation—quiet as it is—may be the most telling signal of all.
